Anthropic publishes a new constitution for Claude

Anthropic has released an updated constitution document that it says directly shapes how Claude is trained and how the model should behave.

Anthropic says it is publishing a new “constitution” for its AI model Claude today, describing how the model should reason and behave.

The announcement points readers to the full text at anthropic.com/constitution.
Anthropic says the constitution is a key input to Claude’s training and directly shapes the model’s intended behavior (even if outputs don’t always match).
The update shifts from a standalone principle list to a longer, rationale-heavy document meant to improve judgment in unfamiliar situations.
Anthropic describes four target traits—helpfulness, guideline compliance, ethical conduct, and safety—and says safety should win conflicts (then ethics, compliance, and helpfulness).
Anthropic says Claude can use the constitution to generate synthetic training data (including conversations and response rankings) for future model training, and is releasing the document under Creative Commons CC0 1.0.